Where the data comes from

Every answer on this site is derived from vendor and project sources, not written by hand. This page lists those sources, the rules applied to them, and the limits. Every entry also carries its own sources in its notes.

Sources, by dataset

• CUDA (21 releases): NVIDIA's CUDA Toolkit release notes for minimum drivers and supported GPU architectures, including NVIDIA's archived notes for older versions.
• PyTorch (18 releases): the files PyTorch actually publishes (download.pytorch.org and PyPI) for which CUDA, ROCm and Python builds exist; PyTorch's own build scripts at each release tag for which GPU architectures each build was compiled for; PyPI package metadata for which build a plain pip install torch gets.
• ROCm (54 releases): AMD's ROCm documentation for each release: the compatibility matrix, and the per-card system requirements, which name each supported GPU and mark deprecated ones.
• TensorFlow (17 releases): TensorFlow's tested build configurations.
• Libraries (6) and inference engines (11 entries): each project's installation docs and build configuration at a specific release tag.
• GPU specs (65 GPUs): NVIDIA, AMD and Intel product pages, datasheets and architecture whitepapers. Where a number is derived (for example, bandwidth from memory speed and bus width), the entry says how. No secondary sources: a spec the vendor doesn't publish is left blank.
• LLM models (39): each model's own configuration on Hugging Face.

The rules the answers use

• NVIDIA binary compatibility. Code built for compute capability X.Y runs on a GPU with the same major version X and a minor version of Y or higher (8.6 code runs on 8.9), never across major versions. PyTorch's own compatibility check uses the same rule.
• Driver requirements. Each CUDA version's full minimum driver, plus NVIDIA's minor version compatibility: CUDA 11.x, 12.x and 13.x builds run on any driver from 450, 525 and 580 respectively, except for code compiled at runtime (such as torch.compile).
• AMD support is per card. A card is officially supported in a ROCm release only if AMD names it there. Cards that only share a supported gfx target are shown as "runs, not officially listed". AMD code doesn't run across gfx targets.
• Native number formats. BF16, FP8 and FP4 support comes from NVIDIA's TensorRT hardware precision table and AMD's ROCm precision-support tables. Formats a vendor marks as emulated don't count as native.
• Consistency checks. The site fails to build if its data contradicts itself, for example if a ROCm release names a card whose gfx target that release doesn't list.

What this site doesn't do

• No performance benchmarks: none are measured here, so none are published.
• PyTorch architecture lists are for Linux x86_64 wheels; Windows wheels are assumed to match, as PyTorch documents for recent releases.
• Compatibility is by version, not by individual driver branch or operating system release.

Use the data

Every dataset is available as JSON: see download the data.

When something is wrong

Use "Report an error on this page" at the bottom of any page. Every correction is recorded in thecorrections log (7 so far), with what the site said before and what it says now.